Evaluation of Quantitative Structure Property Relationship Algorithms for Predicting Plasma Protein Binding in Humans

Area of Science:

• Pharmacokinetics and Drug Metabolism
• Environmental Chemistry
• Computational Chemistry

Background:

• Plasma protein binding, specifically the fraction unbound in plasma (fup), is crucial for predicting compound behavior and exposure using physiologically based pharmacokinetic (PBPK) modeling.
• Quantitative structure-property relationship (QSPR) models offer a method for predicting fup when experimental data are unavailable.
• Existing QSPR models were primarily trained on pharmaceutical compounds, necessitating an evaluation of their accuracy for environmentally relevant substances.

Purpose of the Study:

• To compare the prediction accuracy of three QSPR models (Ingle et al., Watanabe et al., ADMET Predictor) for fraction unbound in plasma (fup).
• To assess the performance of these models for both pharmaceutical and environmentally relevant compounds.
• To identify key chemical descriptors influencing fup prediction accuracy.

Main Methods:

• Calculation of fup values using three distinct QSPR models: Ingle et al., Watanabe et al., and ADMET Predictor.
• Testing the models on a diverse dataset of 818 pharmaceutical and environmentally relevant compounds with fup ranging from 0.01 to 1.
• Analysis of prediction errors, including mean absolute error (MAE) and mean absolute relative prediction error (RPE), across different fup ranges.

Main Results:

• All three QSPR models exhibited a tendency to over-predict fup for highly binding compounds and under-predict for low to moderately binding compounds.
• The Watanabe et al. model demonstrated superior performance for highly binding compounds (0.01 ≤ fup ≤ 0.25), with lower MAE and RPE.
• Ingle et al. and ADMET Predictor showed better accuracy for low to moderately binding compounds.
• Positive polar surface area, number of basic functional groups, and lipophilicity were identified as the most influential chemical descriptors for fup prediction.

Conclusions:

• Prediction of fup using QSPR models is most uncertain for highly binding compounds.
• The performance of QSPR models varies depending on the compound's binding affinity and the specific model used.
• QSPR-derived fup values should be applied with caution in physiologically based pharmacokinetic (PBPK) modeling, particularly for highly binding substances.